CN107220337A - A kind of cross-media retrieval method based on mixing migration network - Google Patents
A kind of cross-media retrieval method based on mixing migration network Download PDFInfo
- Publication number
- CN107220337A CN107220337A CN201710378474.3A CN201710378474A CN107220337A CN 107220337 A CN107220337 A CN 107220337A CN 201710378474 A CN201710378474 A CN 201710378474A CN 107220337 A CN107220337 A CN 107220337A
- Authority
- CN
- China
- Prior art keywords
- media
- mrow
- msub
- network
- msubsup
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/40—Information retrieval; Database structures therefor; File system structures therefor of multimedia data, e.g. slideshows comprising image and additional audio data
- G06F16/43—Querying
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N5/00—Computing arrangements using knowledge-based models
- G06N5/02—Knowledge representation; Symbolic representation
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Data Mining & Analysis (AREA)
- Computational Linguistics (AREA)
- Databases & Information Systems (AREA)
- Multimedia (AREA)
- Evolutionary Computation (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Audiology, Speech & Language Pathology (AREA)
- General Health & Medical Sciences (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The present invention relates to a kind of cross-media retrieval method based on mixing migration network, comprise the following steps:1. setting up monomedia database and across media database, and training set and test set will be divided into across the data in media database.2. using monomedia database and across the data training mixing migration network in the training set of media database, the Unified Characterization for learning different media datas.3. using the mixing migration network trained, the Unified Characterization across the test intensive data of media database is obtained, and then calculate across media similitudes.4. using a kind of medium type in across media test sets as query set, another medium type is retrieved as search library, and final retrieval result is obtained according to similitude.The present invention had both been realized from monomedia to the knowledge migration across media, and the Unified Characterization for being more suitable for cross-media retrieval is generated also by the semantic association for emphasizing aiming field, the accuracy rate of cross-media retrieval is improved.
Description
Technical field
The invention belongs to multimedia retrieval field, and in particular to a kind of cross-media retrieval side based on mixing migration network
Method.
Background technology
With the progress and the development of science and technology of human civilization, the multi-medium data such as image, text, video, audio is fast
Speed increases, and has been increasingly becoming information storage and the principal mode propagated.In this case, cross-media retrieval becomes manually
One of important application of intelligence.Cross-media retrieval is a kind of new retrieval form, can be looked into according to the user of any medium type
Ask, returning has correlation semantic and the different retrieval result of medium type.As user can be used as inquiry inspection by an image
Suo Xiangguan text, or meet by query and search of text the image of its description.Compared with monomedia is retrieved, cross-media retrieval energy
More flexible retrieval experience enough is provided for Internet user, and technical support is provided for the unified management of multi-medium data.So
And, the computer characteristic representation of different media types is often different, constitutes " media wide gap " so that cross-media retrieval face
Face significant challenge.
For the problem, existing main stream approach is Unified Characterization study, and as different media types learn a unification
Semantic space, and then realize cross-media retrieval using modes such as conventional distance metrics.Within this space, the distance of data takes
Certainly in its semantic similarity degree, and it is unrelated with medium type.Prior efforts mainly learn different matchmakers using statistical analysis technique
Linear Mapping matrix of the body to uniform spaces.Such as canonical correlation analysis (canonical correlation analysis, abbreviation
CCA) method can learn the subspace for obtaining maximizing two kinds of media data associations, and then utilize the Unified Characterization of identical dimensional
To represent the data of different media.Rasiwasia et al. is in document " A New Approach to Cross-Modal
High-level semantic mapping method is proposed in Multimedia Retrieval ", learns the unified sky of two kinds of media first with CCA
Between, reuse logistic regression method high-level semantic is obtained in this space and be denoted as final Unified Characterization.This method is in CCA
On the basis of make use of concept classification markup information, therefore the Unified Characterization learnt has more preferable Semantic judgment ability.
Peng et al. is in document " Semi-Supervised Cross-Media Feature Learning with Unified
S is proposed in Patch Graph Regularization "2UPG methods, are combined using same hypergraph and model media
Overall data and semantic primitive, excavate fine-grained across media incidence relations so as to deep, achieve higher retrieval accurate
True rate.
In recent years, deep learning has played important function in multimedia retrieval field, across the media systems based on depth network
One representative learning becomes a study hotspot.As Ngiam " is proposed many in document in Multimodal Deep Learning "
Mode self-encoding encoder method, by considering monomedia reconstructed error and across media incidence relations simultaneously, utilizes a shared coding
Layer obtains Unified Characterization.Srivastava et al. is in document " Learning Representations for Multimodal
Multi-modal depth belief network is proposed in Data with Deep Belief Nets ", passes through two points of an association RBM connection
From depth belief network obtain Unified Characterization.Peng et al. is in document " Cross-media Retrieval by
Proposed in Multimodal Representation Fusion with Deep Networks " across many depth networks of media
Method, by combining a variety of depth network structures, while analyzing in media, the incidence relation between media and carrying out stratification and melt
Close, achieve more preferable effect.
Scale of the actual effect of deep learning method often with training data is closely related, and existing based on depth network
Cross-media retrieval method be faced with the not enough challenge of training data.In monomedia retrieval, researcher is often based on migration and learned
Thought is practised, by the knowledge migration in large-scale dataset (source domain) into the not enough specific tasks (aiming field) of training data, from
And alleviate the problem of training data is not enough.Such as the pre-training model based on ImageNet image data sets is widely used in various
Computer vision problem, significantly improves the effect of model.The abundant semantic knowledge that these extensive monomedia data sets are included
It is not only able to instruct the semantic study of monomedia, can also provides the high-layer semantic information unrelated with medium type to promote across matchmaker
Body semanteme study, has important value for cross-media retrieval.Existing method can only often be instructed using across media data collection
Practice, because training data is not enough and easily causes over-fitting, reduce retrieval effectiveness;Or knowledge is only carried out between media of the same race
Migration, have ignored the knowledge migration between different media so that the transition process of knowledge is not comprehensive enough, limit Unified Characterization study
Effect.
The content of the invention
In view of the shortcomings of the prior art, the present invention proposes a kind of cross-media retrieval method based on mixing migration network,
Monomedia source domain can be realized to the knowledge migration across media object domain in a unified network structure.This method is with source
The medium type being had in domain, aiming field is bridge, and the semantic knowledge in monomedia source domain is moved to across media object simultaneously
In all medium types in domain, and highlight by way of Internet is shared across the media incidence relations in inherence in aiming field,
So that Unified Characterization is more suitable for the cross-media retrieval task on aiming field, retrieval rate is improved.
To achieve the above objectives, the technical solution adopted by the present invention is as follows:
A kind of cross-media retrieval method based on mixing migration network, the Unified Characterization for learning different media datas,
So as to realize cross-media retrieval, comprise the following steps:
(1) set up and include the monomedia database of medium type a kind of, at the same set up include multiple media type across matchmaker
Volume data storehouse, and the data across in media database are divided into training set and test set;
(2) using monomedia database and across the data training mixing migration network in the training set of media database, use
In the Unified Characterization for learning different media datas;
(3) using the mixing migration network trained, the Unified Characterization across the test intensive data of media database is obtained,
And then calculate across media similitudes;
(4) use a kind of medium type in across media test sets as query set, another medium type is used as retrieval
Storehouse, calculates similitude by the use of each data in query set as each data in sample, with search library, is obtained according to similitude
To retrieval result.
Further, the monomedia in above-mentioned a kind of cross-media retrieval method based on mixing migration network, the step (1)
Database includes a kind of medium type;Include at least two medium types across media database, one of which is monomedia data
The medium type in storehouse.
Further, the mixing in above-mentioned a kind of cross-media retrieval method based on mixing migration network, the step (2) is moved
Move network structure, including the shared migration sub-network of media, two parts of the shared association sub-network of Internet.The media are shared to be moved
Sub-network is moved using the shared media in source domain, aiming field as bridge, joint carries out monomedia migration and across media migration so that single
Knowledge in media source domain can be moved to all media in aiming field simultaneously;The shared association sub-network profit of the Internet
The mode shared with Internet so that belong to the close Unified Characterization of across the media data generation of identical semantic classes, so move
Move past journey and taken into full account across the media incidence relations resided in aiming field, improve the retrieval effectiveness of Unified Characterization.
Further, a kind of above-mentioned cross-media retrieval method based on mixing migration network, the similitude of the step (3) takes
Certainly in the distance between Unified Characterization across media data.Distance between Unified Characterization is smaller, then similitude is bigger.
Further, a kind of above-mentioned cross-media retrieval method based on mixing migration network, across the media inspections of the step (4)
Rope mode is, using a kind of media type data in test set as inquiry, another medium type number in retrieval test set
According to correlated results.Similitude is calculated by the similarity calculation method in step (3) and sorted from big to small, is finally examined
Rope the results list.
Effect of the invention is that:Compared with the conventional method, this method can realize monomedia source domain to across media object
The knowledge migration in domain, not only including the knowledge migration between media of the same race, also including the knowledge migration between different media types;
Simultaneously during knowledge migration, across the media semantic association relations that aiming field is contained can be fully excavated.In Unified Characterization
Learning process in, this method is associated using the semantic knowledge of source domain with aiming field across media simultaneously, improves across media inspections
The accuracy rate of rope.
Why this method has foregoing invention effect, and its reason is:Pass through the shared migration sub-network of media and network
The collaborative modeling of the shared association sub-network of layer, while realizing from monomedia to the knowledge migration across media, also by emphasizing
The semantic association generation of aiming field is more suitable for the Unified Characterization of cross-media retrieval.The wherein shared migration sub-network of media is with source domain
Shared media with aiming field are bridge so that monomedia migration can combine progress with across media migration;Internet is shared to close
Connection sub-network then strengthens across media associations by way of layer is shared so that Unified Characterization is more conformed on across media object domain
The semantic association known.Above-mentioned two sub-network constitutes co-ordinative construction end to end, takes full advantage of the semanteme in monomedia source domain
Knowledge promotes the Unified Characterization study on across media object domain, improves the accuracy rate of cross-media retrieval.
Brief description of the drawings
Fig. 1 is the schematic flow sheet of the cross-media retrieval method based on mixing migration network of the present invention.
Fig. 2 is the structural representation of the mixing migration network of the present invention.
Embodiment
The present invention is described in further detail with specific embodiment below in conjunction with the accompanying drawings.
A kind of cross-media retrieval method based on mixing migration network of the present invention, its flow is as shown in figure 1, comprising following
Step:
(1) set up and include the monomedia database of medium type a kind of, at the same set up include multiple media type across matchmaker
Volume data storehouse, and the data across in media database are divided into training set and test set.
In the present embodiment, the medium type that described monomedia database is included is image, is included across media database
Medium type is image, text.For image, spy is used as using the convolutional neural networks structure based on AlexNet in a network
Extractor is levied, this method equally supports other to be used for the convolutional neural networks structure of image characteristics extraction, such as VGG-19;For
Text, feature is used as using word frequency vector.This method equally supports the media and feature of other species, need to correspond to and replace corresponding
Feature extraction network or method.
Use DsRepresent monomedia database (being used as source domain), Ds={ Is}.WhereinRepresent with mark
Image,For r-th of image, its classification marked isM represents the amount of images in monomedia database.
It is expressed as across media database (being used as aiming field)Picture/text wherein with mark
To for { It,Tt, the picture/text not marked to forWith monomedia class database seemingly, define WithWhereinFor p-th of image, if it is band labeled data, mark
The classification of note is For q-th of band mark text, if it is band labeled data, the classification marked isnlRepresent band mark
The picture/text of note is to quantity, and n is represented with mark and the picture/text not marked to total amount.It regard the data with mark as instruction
Practice collection, the data not marked are as test set.
The target of study is by DsIn the knowledge migration that includes to DtIn, and beLearn Unified CharacterizationWithAnd then cross-media retrieval task is carried out using Unified Characterization, whereinFor p-th of figure
The Unified Characterization of picture,For the Unified Characterization of q-th of text.
(2) using monomedia database and across the data training mixing migration network in the training set of media database, use
In the Unified Characterization for learning different media datas.
Network structure constructed by the step is as shown in Fig. 2 including two sub-networks:The shared migration sub-network of media and net
The shared association sub-network of network layers.Conv1~conv5 is convolutional layer in figure, and fc6~fc10 is full articulamentum, and Softmax is classification
Device.In the training process, whole network has the input of three tunnels:Source domain image, target area image, aiming field text.For media
Shared migration sub-network, three tunnels input is handled respectively through two layers of fully-connected network, and joint considers monomedia migration and across media
Two aspects are migrated, so as to realize from monomedia source domain to the knowledge migration across media object domain.
Wherein, monomedia migration occurs between source domain image and target area image two-way network.In the present embodiment, adopt
Adapt to (feature adaption) mode with feature to carry out, target is the maximum mean difference for minimizing view data in two domains
Different (maximum mean discrepancy, abbreviation MMD).The different distributions that a and b is media data of the same race on two domains are made,
Then the MMD of the two is in regeneration Hilbert space HkIn quadratic form be defined as follows:
Wherein, isAnd itSource domain, the input picture of aiming field are represented respectively,Represent a certain layer of correspondence in depth network
Output, EaAnd EbExpectation in expression source domain, aiming field.Then, monomedia migration loss is built as follows:
Wherein, l6With l7For corresponding full articulamentum in source domain image, target area image two-way network, i.e. fc6- in Fig. 2
S/fc6-i and fc7-s/fc-7i.In addition, source domain image network is also required in IsOn be finely adjusted to excavate source domain and aiming field
Semantic association information.Source domain supervision loss item is set up for this:
Wherein m represents the amount of images for having mark in source domain,Item is lost for softmax, is defined as follows:
Wherein θ is network parameter, and y is data x label, and c is all possible number of labels of x.1 { y=j } is instruction letter
Number, 1 is returned if y=j, is otherwise 0.
Across media migration generation is between target area image and aiming field text.Think each picture/text to phase
The high-level semantic of pass, therefore it is similar to realize across media knowledge migrations to export their Internet;Make each picture/text pair
ForDefine across media differences as follows:
It is as follows that across media migration loss item is set up accordingly:
In the training process of network, network migrates monomedia as bridge and across media migration target area image all the way
Join together, realize the knowledge migration of monomedia source domain all medium types into across media object domain.
For the shared association sub-network of Internet, it is inputted comprising target area image and aiming field text two-way, and two-way
Two full articulamentums are shared in input, form Internet shared structure.To maintain across media semantic association information, using mentioned above
Output of the softmax loss items to every kind of media be modeled, then across media associations lose items and are:
Wherein, fsLost for foregoing softmax,WithThe respectively supervision of image, text
Bound term.Because whole network constitutes structure end to end, the shared migration sub-network of media, the shared association sub-network of Internet can
To carry out joint training by back-propagation algorithm, so that transition process maintains across the media associations in aiming field, more
It is adapted to cross-media retrieval task.
(3) using the mixing migration network trained, the Unified Characterization across the test intensive data of media database is obtained,
And then calculating is across the similitude of media data.
Because mixing is migrated, network is trained in (2) to be finished, and the network directly can be made with a kind of data of media
For input, its Unified Characterization is obtained, and input is used as without monomedia source.Unified Characterization is depended on across the similitude of media data
The distance between.Distance between Unified Characterization is smaller, then similitude is bigger.In the present embodiment, by calculating two different matchmakers
COS distance between the Unified Characterization of volume data reflects the similitude of the two.The present invention equally supports other kinds of distance
Measurement, such as Euclidean distance.
(4) a kind of medium type in step (1) in across media test sets is used as query set, another medium type
It is used as search library.By the use of each data in query set as sample, according to the similarity calculation method in step (3), with inspection
Each data in Suo Ku calculate similitude and sorted from big to small, return to retrieval result list.
Following experiments result shows that compared with the conventional method, the present invention migrates the cross-media retrieval side of network based on mixing
Method, can obtain higher retrieval rate.
In the present embodiment, there are about 1,200,000 with mark image for conventional as the monomedia database of source domain
ImageNet data set subset ILSVRC2012, across media database then be across the media data collection of Wikipedia.Wikipedia
Data set includes 2866 picture/texts pair, is divided into 10 classifications, each picture/text is to corresponding relation.The number
According to collection by document " A New Approach to Cross-Modal Multimedia Retrieval " (authors
N.Rasiwasia,J.Pereira,E.Coviello,G.Doyle,G.Lanckriet,R.Levy and
N.Vasconcelos) propose, be one of the most frequently used data set in cross-media retrieval.According to document " Cross-modal
Drawing in Retrieval with Correspondence Autoencoder " (author F.Feng, X.Wang, and R.Li)
The mode of dividing, Wikipedia data sets is divided into the training set for including 2173 picture/texts pair, 462 picture/texts pair
Test set, and 231 picture/texts pair checking collection.Wherein training set, test set directly correspond to step (1) in across
Media database training set, test set, checking collection are used for control methods regulation parameter, and the present invention need not verify collection conduct
Input.Following 3 kinds of methods are tested as Experimental comparison:
Existing method one:Document " Learning Cross-media Joint Representation with Sparse
Unified representation in and Semi-supervised Regularization " (author X.Zhai, Y.Peng and J.Xiao)
Learn (JRL) method, be that it learns Unified Characterization mapping square by analyzing the incidence relation and high-layer semantic information of different media
Battle array;
Existing method two:Document " Cross-modal Retrieval with Correspondence
Consistent self-encoding encoder (Corr-AE) method in Autoencoder " (author F.Feng, X.Wang, and R.Li), builds deep
Self-encoding encoder network is spent, using two kinds of media type datas as input, while across the media incidence relations of modeling and reconstructed error, are being compiled
Code layer obtains the Unified Characterization of different media;
Existing method three:Document " Cross-modal retrieval with CNN visual features:A new
The method proposed in baseline " (author Y.Wei, Y.Zhao, C.Lu, S.Wei, L.Liu, Z.Zhu, and S.Yan), to figure
Carry out high-level semantic study as, text builds a road network respectively, carried out using semantic classes probability vector as Unified Characterization across
Media retrieval.Wherein image is trained using model based on ILSVRC2012 pre-training convolutional neural networks all the way.
The present invention:The method of the present embodiment.
In this experiment, Average Accuracy average (the mean average commonly used using information retrieval field
Precision, abbreviation MAP) accuracy of cross-media retrieval is evaluated and tested as index, the bigger explanation cross-media retrieval of MAP value
As a result it is better.
The Experimental results show of the present invention of table 1.
Method | Image retrieval text | Text retrieval image | It is average |
Existing method one | 0.408 | 0.353 | 0.381 |
Existing method two | 0.373 | 0.357 | 0.365 |
Existing method three | 0.458 | 0.345 | 0.402 |
The present invention | 0.508 | 0.432 | 0.470 |
As it can be seen from table 1 the present invention is in image retrieval text, two tasks of text retrieval image, existing side is contrasted
Method all achieves the raising of accuracy rate.The study of method one Linear Mapping obtains the Unified Characterization of different media datas, it is difficult to fully
Model the complexity association across media data.Control methods two and method three, the method that they are all based on deep learning, but method
Two paired associations for make use of across media data, its high-layer semantic information of modeling not yet in effect;Method three employs transfer learning
Thought, but the migration of monomedia is pertained only to, the knowledge migration process across media is ignored, therefore fail to make full use of monomedia source
Abundant semantic knowledge in domain.The present invention is built by the shared migration sub-network of media with the shared collaboration for associating sub-network of Internet
Mould, had both been realized from monomedia to the knowledge migration across media, was more fitted also by the semantic association of aiming field is emphasized with generating
Close the Unified Characterization of cross-media retrieval.Two sub-networks constitute co-ordinative construction end to end, make full use of in monomedia source domain
Semantic knowledge promotes the Unified Characterization study on across media object domain, improves the accuracy rate of cross-media retrieval.
In addition to above example, the specific steps of the present invention program can also use other embodiment, such as in step (1)
In the monomedia database of other medium types (such as text) can be used, and be not limited to two kinds of image, text across media database
Medium type;It is that each medium type is used in other kinds of character representation, or increase, reduction network structure in step (2)
The number of convolutional layer and full articulamentum, softmax graders etc. are substituted using other graders, fall within the deformation side of the present invention
Formula.
Obviously, those skilled in the art can carry out the essence of various changes and modification without departing from the present invention to the present invention
God and scope.So, if these modifications and variations of the present invention belong to the scope of the claims in the present invention and its equivalent technologies
Within, then the present invention is also intended to comprising including these changes and modification.
Claims (9)
1. a kind of cross-media retrieval method based on mixing migration network, comprises the following steps:
(1) the monomedia database for including a kind of medium type is set up, while setting up across the media numbers for including multiple media type
It is divided into training set and test set according to storehouse, and by the data across in media database;
(2) using monomedia database and across the data training mixing migration network in the training set of media database, for learning
Practise the Unified Characterization of different media datas;
(3) using the mixing migration network trained, the Unified Characterization across the test intensive data of media database is obtained, and then
Calculate the similitude across media data;
(4) a kind of medium type in across media test sets is used as query set, and another medium type is used as search library, profit
Similitude is calculated as each data in sample, with search library with each data in query set, is examined according to similitude
Hitch fruit.
2. the method as described in claim 1, it is characterised in that the monomedia database in step (1) includes a kind of media class
Type;
Include two kinds of medium types across media database, one of which is the medium type of monomedia database.
3. the method as described in claim 1, it is characterised in that the mixing migration network structure in step (2), including media are common
Enjoy that migration sub-network and Internet are shared associate sub-network, the shared sub-network that migrates of the media is with being total in source domain, aiming field
It is bridge to have media, and joint carries out monomedia migration and across media migration so that the knowledge in monomedia source domain can be by simultaneously
Move to all media in aiming field;The shared association sub-network of the Internet utilizes the mode that Internet is shared so that category
In the close Unified Characterization of across the media data generation of identical semantic classes, take into full account transition process and reside in aiming field
Across media incidence relations, to improve the retrieval effectiveness of Unified Characterization.
4. method as claimed in claim 3, it is characterised in that the shared migration sub-network of media in step (2) has three tunnels
Input:Source domain image, target area image, aiming field text, three tunnels input are handled respectively through two layers of fully-connected network;In training
During joint consider monomedia migration and realize from monomedia source domain to across media object domain across two aspects of media migration
Knowledge migration;The monomedia migration occurs between source domain image and target area image two-way network, described to be moved across media
Move and occur between target area image and aiming field text.
5. method as claimed in claim 4, it is characterised in that the monomedia migration is carried out using feature accommodation mode, fixed
View data a and b maximum mean difference are in regeneration Hilbert space H in adopted two domainskIn quadratic form be:
OrderFor the output of a certain layer of correspondence in depth network, monomedia migration loss is built as follows:
<mrow>
<msub>
<mi>Loss</mi>
<mrow>
<mi>S</mi>
<mi>i</mi>
<mi>n</mi>
<mi>g</mi>
<mi>l</mi>
<mi>e</mi>
</mrow>
</msub>
<mo>=</mo>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>l</mi>
<mo>=</mo>
<msub>
<mi>l</mi>
<mn>6</mn>
</msub>
</mrow>
<msub>
<mi>l</mi>
<mn>7</mn>
</msub>
</munderover>
<msubsup>
<mi>d</mi>
<mi>k</mi>
<mn>2</mn>
</msubsup>
<mrow>
<mo>(</mo>
<msup>
<mi>I</mi>
<mi>s</mi>
</msup>
<mo>,</mo>
<msup>
<mi>I</mi>
<mi>t</mi>
</msup>
<mo>)</mo>
</mrow>
<mo>,</mo>
</mrow>
Wherein, l6With l7For corresponding full articulamentum in source domain image, target area image two-way network;In addition, source domain image network
It is also required in IsOn be finely adjusted to excavate the semantic association information of source domain and aiming field, set up for this source domain supervision loss item:
<mrow>
<msub>
<mi>Loss</mi>
<mrow>
<mi>S</mi>
<mi>o</mi>
<mi>u</mi>
<mi>r</mi>
<mi>c</mi>
<mi>e</mi>
</mrow>
</msub>
<mo>=</mo>
<mfrac>
<mn>1</mn>
<mi>m</mi>
</mfrac>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>r</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>m</mi>
</munderover>
<msub>
<mi>f</mi>
<mi>s</mi>
</msub>
<mrow>
<mo>(</mo>
<msubsup>
<mi>i</mi>
<mi>r</mi>
<mi>s</mi>
</msubsup>
<mo>,</mo>
<msubsup>
<mi>y</mi>
<mi>r</mi>
<mi>s</mi>
</msubsup>
<mo>,</mo>
<msup>
<mi>&theta;</mi>
<mi>s</mi>
</msup>
<mo>)</mo>
</mrow>
<mo>,</mo>
</mrow>
WhereinItem is lost for softmax, is defined as follows:
<mrow>
<msub>
<mi>f</mi>
<mi>s</mi>
</msub>
<mrow>
<mo>(</mo>
<mi>x</mi>
<mo>,</mo>
<mi>y</mi>
<mo>,</mo>
<mi>&theta;</mi>
<mo>)</mo>
</mrow>
<mo>=</mo>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>j</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>c</mi>
</munderover>
<mn>1</mn>
<mo>{</mo>
<mi>y</mi>
<mo>=</mo>
<mi>j</mi>
<mo>}</mo>
<mfrac>
<msup>
<mi>e</mi>
<mrow>
<msub>
<mi>&theta;</mi>
<mi>j</mi>
</msub>
<mi>x</mi>
</mrow>
</msup>
<mrow>
<msubsup>
<mi>&Sigma;</mi>
<mrow>
<mi>l</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>c</mi>
</msubsup>
<msup>
<mi>e</mi>
<mrow>
<msub>
<mi>&theta;</mi>
<mi>l</mi>
</msub>
<mi>x</mi>
</mrow>
</msup>
</mrow>
</mfrac>
<mo>,</mo>
</mrow>
Wherein θ is network parameter, and y is data x label, and c is all possible number of labels of x;1 { y=j } is indicator function,
1 is returned if y=j, is otherwise 0.
6. method as claimed in claim 4, it is characterised in that it is described across each picture/text in media migration to phase
The high-level semantic of pass, therefore make their Internet that there is similar output to realize across media knowledge migrations;Make each image/
Text to for, define across media differences as follows:
It is as follows that across media migration loss item is set up accordingly:
<mrow>
<msub>
<mi>Loss</mi>
<mrow>
<mi>C</mi>
<mi>r</mi>
<mi>o</mi>
<mi>s</mi>
<mi>s</mi>
</mrow>
</msub>
<mo>=</mo>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>l</mi>
<mo>=</mo>
<msub>
<mi>l</mi>
<mn>6</mn>
</msub>
</mrow>
<msub>
<mi>l</mi>
<mn>7</mn>
</msub>
</munderover>
<munderover>
<mi>&Sigma;</mi>
<mrow>
<mi>p</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<msub>
<mi>n</mi>
<mi>l</mi>
</msub>
</munderover>
<msubsup>
<mi>d</mi>
<mi>c</mi>
<mn>2</mn>
</msubsup>
<mrow>
<mo>(</mo>
<msubsup>
<mi>i</mi>
<mi>p</mi>
<mi>t</mi>
</msubsup>
<mo>,</mo>
<msubsup>
<mi>t</mi>
<mi>p</mi>
<mi>t</mi>
</msubsup>
<mo>)</mo>
</mrow>
<mo>.</mo>
</mrow>
7. method as claimed in claim 3, it is characterised in that the shared association sub-network of Internet in step (2) has mesh
Area image and the input of aiming field text two-way are marked, and two full articulamentums are shared in two-way input, form Internet shared structure;For
Across media semantic association information are maintained, losing output of the item to every kind of media using softmax is modeled, then across media associations
Losing item is:
<mrow>
<msub>
<mi>Loss</mi>
<mrow>
<mi>C</mi>
<mi>o</mi>
<mi>r</mi>
<mi>r</mi>
<mi>e</mi>
<mi>l</mi>
<mi>a</mi>
<mi>t</mi>
<mi>i</mi>
<mi>o</mi>
<mi>n</mi>
</mrow>
</msub>
<mo>=</mo>
<mfrac>
<mn>1</mn>
<msub>
<mi>n</mi>
<mi>l</mi>
</msub>
</mfrac>
<munderover>
<mi>&Sigma;</mi>
<mrow>
<mi>p</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<msub>
<mi>n</mi>
<mi>l</mi>
</msub>
</munderover>
<mrow>
<mo>(</mo>
<msub>
<mi>f</mi>
<mi>s</mi>
</msub>
<mo>(</mo>
<mrow>
<msubsup>
<mi>i</mi>
<mi>p</mi>
<mi>t</mi>
</msubsup>
<mo>,</mo>
<msubsup>
<mi>y</mi>
<mi>p</mi>
<mi>t</mi>
</msubsup>
<mo>,</mo>
<msup>
<mi>&theta;</mi>
<mi>t</mi>
</msup>
</mrow>
<mo>)</mo>
<mo>+</mo>
<msub>
<mi>f</mi>
<mi>s</mi>
</msub>
<mo>(</mo>
<mrow>
<msubsup>
<mi>t</mi>
<mi>p</mi>
<mi>t</mi>
</msubsup>
<mo>,</mo>
<msubsup>
<mi>y</mi>
<mi>p</mi>
<mi>t</mi>
</msubsup>
<mo>,</mo>
<msup>
<mi>&theta;</mi>
<mi>t</mi>
</msup>
</mrow>
<mo>)</mo>
<mo>)</mo>
</mrow>
<mo>,</mo>
</mrow>
Wherein, fsLost for softmax,WithThe respectively supervision bound term of image, text.
8. the method as described in claim 1, it is characterised in that in across the media similarity calculation methods of step (3), across media
The similitude of data depends on the distance between Unified Characterization, and the distance between Unified Characterization is smaller, then similitude is bigger.
9. the method as described in claim 1, it is characterised in that the cross-media retrieval in step (4) refers to, from step (1)
Test set in take out the data of any one media as inquiry sample, it is right according to the similarity calculation method in step (3)
All data of another media in test set carry out Similarity measures, and are sorted from big to small according to similitude, return to inspection
Rope the results list.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710378474.3A CN107220337B (en) | 2017-05-25 | 2017-05-25 | Cross-media retrieval method based on hybrid migration network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710378474.3A CN107220337B (en) | 2017-05-25 | 2017-05-25 | Cross-media retrieval method based on hybrid migration network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN107220337A true CN107220337A (en) | 2017-09-29 |
CN107220337B CN107220337B (en) | 2020-12-22 |
Family
ID=59944577
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710378474.3A Active CN107220337B (en) | 2017-05-25 | 2017-05-25 | Cross-media retrieval method based on hybrid migration network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107220337B (en) |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109670071A (en) * | 2018-10-22 | 2019-04-23 | 北京大学 | A kind of across the media Hash search methods and system of the guidance of serializing multiple features |
CN109800677A (en) * | 2018-12-29 | 2019-05-24 | 西安交通大学 | A kind of cross-platform palm grain identification method |
CN110457982A (en) * | 2018-12-28 | 2019-11-15 | 中国科学院合肥物质科学研究院 | A kind of crop disease image-recognizing method based on feature transfer learning |
CN110647904A (en) * | 2019-08-01 | 2020-01-03 | 中国科学院信息工程研究所 | Cross-modal retrieval method and system based on unmarked data migration |
CN110889505A (en) * | 2019-11-18 | 2020-03-17 | 北京大学 | Cross-media comprehensive reasoning method and system for matching image-text sequences |
CN111316290A (en) * | 2017-11-03 | 2020-06-19 | 通用电气公司 | System and method for interactive representation learning migration through deep learning of feature ontologies |
WO2020134409A1 (en) * | 2018-12-28 | 2020-07-02 | 深圳光启空间技术有限公司 | Cross-domain face recognition algorithm, storage medium, and processor |
CN111708745A (en) * | 2020-06-18 | 2020-09-25 | 全球能源互联网研究院有限公司 | Cross-media data sharing representation method and user behavior analysis method and system |
CN111782833A (en) * | 2020-06-09 | 2020-10-16 | 南京理工大学 | Fine-grained cross-media retrieval method based on multi-model network |
Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US7992215B2 (en) * | 2002-12-11 | 2011-08-02 | Trio Systems, Llc | Annotation system for creating and retrieving media and methods relating to same |
CN104166684A (en) * | 2014-07-24 | 2014-11-26 | 北京大学 | Cross-media retrieval method based on uniform sparse representation |
CN104239554A (en) * | 2014-09-24 | 2014-12-24 | 南开大学 | Cross-domain and cross-category news commentary emotion prediction method |
CN103049526B (en) * | 2012-12-20 | 2015-08-05 | 中国科学院自动化研究所 | Based on the cross-media retrieval method of double space study |
US20160071024A1 (en) * | 2014-02-25 | 2016-03-10 | Sri International | Dynamic hybrid models for multimodal analysis |
CN105718532A (en) * | 2016-01-15 | 2016-06-29 | 北京大学 | Cross-media sequencing method based on multi-depth network structure |
CN106202514A (en) * | 2016-07-21 | 2016-12-07 | 北京邮电大学 | Accident based on Agent is across the search method of media information and system |
CN106649434A (en) * | 2016-09-06 | 2017-05-10 | 北京蓝色光标品牌管理顾问股份有限公司 | Cross-domain knowledge transfer tag embedding method and apparatus |
CN106650756A (en) * | 2016-12-28 | 2017-05-10 | 广东顺德中山大学卡内基梅隆大学国际联合研究院 | Image text description method based on knowledge transfer multi-modal recurrent neural network |
-
2017
- 2017-05-25 CN CN201710378474.3A patent/CN107220337B/en active Active
Patent Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US7992215B2 (en) * | 2002-12-11 | 2011-08-02 | Trio Systems, Llc | Annotation system for creating and retrieving media and methods relating to same |
CN103049526B (en) * | 2012-12-20 | 2015-08-05 | 中国科学院自动化研究所 | Based on the cross-media retrieval method of double space study |
US20160071024A1 (en) * | 2014-02-25 | 2016-03-10 | Sri International | Dynamic hybrid models for multimodal analysis |
CN104166684A (en) * | 2014-07-24 | 2014-11-26 | 北京大学 | Cross-media retrieval method based on uniform sparse representation |
CN104239554A (en) * | 2014-09-24 | 2014-12-24 | 南开大学 | Cross-domain and cross-category news commentary emotion prediction method |
CN105718532A (en) * | 2016-01-15 | 2016-06-29 | 北京大学 | Cross-media sequencing method based on multi-depth network structure |
CN106202514A (en) * | 2016-07-21 | 2016-12-07 | 北京邮电大学 | Accident based on Agent is across the search method of media information and system |
CN106649434A (en) * | 2016-09-06 | 2017-05-10 | 北京蓝色光标品牌管理顾问股份有限公司 | Cross-domain knowledge transfer tag embedding method and apparatus |
CN106650756A (en) * | 2016-12-28 | 2017-05-10 | 广东顺德中山大学卡内基梅隆大学国际联合研究院 | Image text description method based on knowledge transfer multi-modal recurrent neural network |
Non-Patent Citations (1)
Title |
---|
YUXIN PENG等: ""Cross-media similarity metric learning with unified deep networks"", 《MULTIMEDIA TOOLS AND APPLICATIONS》 * |
Cited By (17)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111316290A (en) * | 2017-11-03 | 2020-06-19 | 通用电气公司 | System and method for interactive representation learning migration through deep learning of feature ontologies |
CN111316290B (en) * | 2017-11-03 | 2024-01-12 | 通用电气公司 | System and method for interactive representation learning migration through deep learning of feature ontologies |
CN109670071A (en) * | 2018-10-22 | 2019-04-23 | 北京大学 | A kind of across the media Hash search methods and system of the guidance of serializing multiple features |
CN109670071B (en) * | 2018-10-22 | 2021-10-08 | 北京大学 | Serialized multi-feature guided cross-media Hash retrieval method and system |
CN110457982A (en) * | 2018-12-28 | 2019-11-15 | 中国科学院合肥物质科学研究院 | A kind of crop disease image-recognizing method based on feature transfer learning |
WO2020134409A1 (en) * | 2018-12-28 | 2020-07-02 | 深圳光启空间技术有限公司 | Cross-domain face recognition algorithm, storage medium, and processor |
CN110457982B (en) * | 2018-12-28 | 2023-04-11 | 中国科学院合肥物质科学研究院 | Crop disease image identification method based on feature migration learning |
CN109800677A (en) * | 2018-12-29 | 2019-05-24 | 西安交通大学 | A kind of cross-platform palm grain identification method |
CN109800677B (en) * | 2018-12-29 | 2021-11-02 | 掌握科技无锡有限公司 | Cross-platform palm print identification method |
CN110647904B (en) * | 2019-08-01 | 2022-09-23 | 中国科学院信息工程研究所 | Cross-modal retrieval method and system based on unmarked data migration |
CN110647904A (en) * | 2019-08-01 | 2020-01-03 | 中国科学院信息工程研究所 | Cross-modal retrieval method and system based on unmarked data migration |
CN110889505A (en) * | 2019-11-18 | 2020-03-17 | 北京大学 | Cross-media comprehensive reasoning method and system for matching image-text sequences |
CN110889505B (en) * | 2019-11-18 | 2023-05-02 | 北京大学 | Cross-media comprehensive reasoning method and system for image-text sequence matching |
CN111782833A (en) * | 2020-06-09 | 2020-10-16 | 南京理工大学 | Fine-grained cross-media retrieval method based on multi-model network |
CN111782833B (en) * | 2020-06-09 | 2023-12-19 | 南京理工大学 | Fine granularity cross-media retrieval method based on multi-model network |
CN111708745B (en) * | 2020-06-18 | 2023-04-21 | 全球能源互联网研究院有限公司 | Cross-media data sharing representation method and user behavior analysis method and system |
CN111708745A (en) * | 2020-06-18 | 2020-09-25 | 全球能源互联网研究院有限公司 | Cross-media data sharing representation method and user behavior analysis method and system |
Also Published As
Publication number | Publication date |
---|---|
CN107220337B (en) | 2020-12-22 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107220337A (en) | A kind of cross-media retrieval method based on mixing migration network | |
CN107562812B (en) | Cross-modal similarity learning method based on specific modal semantic space modeling | |
Cong et al. | Gan memory with no forgetting | |
CN112214610B (en) | Entity relationship joint extraction method based on span and knowledge enhancement | |
Xu et al. | Augmenting strong supervision using web data for fine-grained categorization | |
Patterson et al. | Coco attributes: Attributes for people, animals, and objects | |
Masko et al. | The impact of imbalanced training data for convolutional neural networks | |
CN111488734A (en) | Emotional feature representation learning system and method based on global interaction and syntactic dependency | |
CN111753098A (en) | Teaching method and system based on cross-media dynamic knowledge graph | |
CN106095893A (en) | A kind of cross-media retrieval method | |
CN112711953A (en) | Text multi-label classification method and system based on attention mechanism and GCN | |
CN105718532B (en) | A kind of across media sort methods based on more depth network structures | |
CN105512209A (en) | Biomedicine event trigger word identification method based on characteristic automatic learning | |
CN112906770A (en) | Cross-modal fusion-based deep clustering method and system | |
CN108388639B (en) | Cross-media retrieval method based on subspace learning and semi-supervised regularization | |
CN110751216A (en) | Judgment document industry classification method based on improved convolutional neural network | |
Tan | Information Analysis of Advanced Mathematics Education‐Adaptive Algorithm Based on Big Data | |
Xu et al. | Large-margin multi-view Gaussian process for image classification | |
Zhang et al. | A self-supervised deep learning framework for unsupervised few-shot learning and clustering | |
Lonij et al. | Open-world visual recognition using knowledge graphs | |
Qu et al. | The rise of ai language pathologists: Exploring two-level prompt learning for few-shot weakly-supervised whole slide image classification | |
CN114021584A (en) | Knowledge representation learning method based on graph convolution network and translation model | |
Zhang et al. | A YOLOv7 incorporating the Adan optimizer based corn pests identification method | |
CN114036308A (en) | Knowledge graph representation method based on graph attention neural network | |
CN108228779A (en) | A kind of result prediction method based on Learning Community's dialogue stream |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |